frankzhangyk / ct-wgan_vgg_tensorflow

Low Dose CT Image Denoising Using a Generative Adversarial Network with Wasserstein Distance and Perceptual Loss

WGAN_VGG_tensorflow

Low Dose CT Image Denoising Using a Generative Adversarial Network with Wasserstein Distance and Perceptual Loss

• WGAN_VGG

• paper : https://arxiv.org/pdf/1708.00961.pdf
• original code:
  • vgg : https://github.com/machrisaa/tensorflow-vgg
  • WGAN : https://github.com/jiamings/wgan

I/O (DICOM file -> .npy)

• Input data Directory
  • DICOM file extension = ['.IMA', '.dcm']

$ os.path.join(dcm_path, patient_no, [LDCT_path|NDCT_path], '*.' + extension)

Ntwork architecture

• Generator

• 8 conv layers (with relu)
• kernel size : 3*3
• filters
  • first_7 : 32
  • last : 1

• discriminator

• str :
  • 6 conv layer(with leaky relu)
    -> fully connected : 1024 (with leaky relu)
    -> fully connected : 1 (cross entropy)
• filters:
  • first_2 : 64
  • mid_2 : 128
  • last_2 : 256
  • kernel size : 3*3

• Perceptual loss

• pre-trained VGG net
• feed : ground truth (generated image)
• loss: L2 (ground truth)
• gradient : only generator

Training detail

• mini batch size L : 128
• opt : Adam(alpha = 1e-5, beta1 = 0.5, beta2 = 0.9)
• discriminator iter : 4
• lambda(WGAN weight penalty) : 10
• lambda1(VGG weight) : 0.1, lambda2 : 0.1

Different

[X] epoch : 100 -> iteration
[X] remove mostly air images -> no remove (Hounsfield unit scale -> normalize 0 ~ 1)

Main file(main.py) Parameters

• Directory

• dcm_path : dicom file directory
• LDCT_path : LDCT image folder name
• NDCT_path : NDCT image folder name
• test_patient_no : test patient id list(p_id1,p_id2...) (train patient id : (patient id list - test patient id list)
• checkpoint_dir : save directory - trained model
• test_npy_save_dir : save directory - test numpy file
• pretrained_vgg : pretrained vggnet directory

• Image info

• patch_size : patch size
• whole_size : whole size
• img_channel : image channel
• img_vmax : max value
• img_vmin : min value

• Train/Test

• model : red_cnn, wgan_vgg, cyclegan (for image preprocessing)
• phase : train | test

• others

• is_mayo : summary ROI sample1,2
• save_freq : save a model every save_freq (iterations)
• print_freq : print_freq (iterations)
• continue_train : load the latest model: true, false
• gpu_no : visible devices(gpu no)

• Training detail

• num_iter : iterations (default = 200000)
• alpha : learning rate (default=1e-5)
• batch_size : batch size (default=128)
• d_iters : discriminator iteration (default=4)
• lambda_ : Gradient penalty term weight (default=10)
• lambda_1 : Perceptual loss weight (in WGAN_VGG network) (default=0.1)
• #lambda_2 : MSE loss weight(in WGAN_VGG network) (default=0.1, not used)
• beta1 : Adam optimizer parameter (default=0.5)
• beta2 : Adam optimizer parameter (default=0.9)

Run

• train

python main.py

• test

python main.py --phase=test

pretrained vgg

• https://github.com/machrisaa/tensorflow-vgg
• https://mega.nz/#!xZ8glS6J!MAnE91ND_WyfZ_8mvkuSa2YcA7q-1ehfSm-Q1fxOvvs